CO2 geological sequestration inversion simulation method based on deep learning parameterization strategy

Abstract

The invention discloses a CO2 geological sequestration inversion simulation method based on a deep learning parameterization strategy, and the method comprises the steps: carrying out the parameterization dimension reduction representation of a complex permeability field of a CO2 geological sequestration storage layer through a deep learning method, and representing the distribution characteristics of a heterogeneous permeability field through a group of latent vectors in low-dimensional normal distribution, and the number of to-be-estimated parameters in CO2 geological sequestration numerical simulation is reduced. According to the method, a parameterization method based on deep learning and a multi-data assimilation iteration set smoothing algorithm are combined, the heterogeneous permeability field of a complex reservoir stratum can be efficiently estimated only by cyclically updating the low-dimensional latent vector during inversion, and then the simulation and prediction precision of a numerical model on the CO2 geological storage process is improved. According to the method, by effectively estimating the complex permeability distribution of the CO2 geological storage layer, the simulation description precision of the CO2 migration and transformation process is improved, and technical support can be provided for optimized operation and risk management of a CO2 geological storage project.

Description

Technical Field

[0001] The present invention relates to the cross-technical field of hydrogeological numerical simulation, carbon emission reduction and deep learning, and specifically to a CO2 geological storage inversion simulation method based on a deep learning parameterization strategy. Background Art

[0002] CO2 geological storage (GCS) is a technical means of injecting CO2 into deep aquifers for permanent storage, which can effectively reduce greenhouse gas emissions and achieve the goal of "carbon neutrality". After injection, CO2 exists in the reservoir as a supercritical phase, forming a CO2-water multiphase flow system. Since the CO2 geological storage process usually involves a spatial scale of hundreds of kilometers and a time scale of hundreds of years, and the deep geological environment is often invisible, numerical simulation methods have become an important tool for studying the migration and fate of CO2 in the formation after injection. Reliable numerical models need to accurately characterize the hydrogeological parameters of the reservoir (such as permeability, etc.), but due to the heterogeneity of the medium itself and the lack of observational data, it is usually not feasible to directly measure the hydrogeological parameters of the reservoir. Therefore, the unknown parameter field is often estimated by solving the inversion problem, that is, given the historical observation data of the relevant state variables, the hydrogeological parameter values ​​are continuously adjusted through the inversion algorithm to reduce the fitting error between the numerical model simulation value and the historical observation data, and reproduce the actual CO2 plume migration process. Therefore, inversion simulation provides important decision-making references for optimized operation and risk management of GCS projects.

[0003] However, commonly used set-based inversion algorithms usually assume that the hydrogeological parameter field to be estimated (such as the permeability field) is Gaussian distributed. However, many actual reservoirs usually contain multiple sedimentary facies with different properties, and their hydrogeological parameters show significant non-Gaussian distribution characteristics, which makes the commonly used set-based inversion algorithms difficult to apply. In addition, a large number of parameters to be estimated will bring a huge computational burden to the inversion calculation. These challenges still have many limitations in the practical application of set-based inversion algorithms. In recent years, the deep learning parameterization (DLP) method has received increasing attention in dealing with non-Gaussian parameter inversion. Combining it with the traditional set-based inversion algorithm is a solution worth exploring. It is also worth noting that due to limited training data, the output results of the deep learning parameterization method itself also have certain uncertainties, and the commonly used set-based inversion algorithm often ignores this uncertainty, which leads to unreliable inversion simulation results. Summary of the invention

[0004] The purpose of the present invention is to provide a CO2 geological storage inversion simulation method based on a deep learning parameterization strategy. The method first constructs a deep learning parameterization model to characterize the non-Gaussian distribution characteristics of the reservoir permeability field with a set of low-dimensional normally distributed latent vectors, while greatly reducing the number of parameters to be estimated. Then, combined with a multiple data assimilation iterative ensemble smoothing algorithm, a CO2 geological storage coupled inversion simulation framework is constructed. During inversion, only the low-dimensional latent vectors need to be cyclically updated to efficiently estimate the complex heterogeneous permeability field of the reservoir, thereby improving the simulation prediction accuracy of the numerical model for the CO2 migration and transformation process. The error term of the standard deviation of the output results of the deep learning parameterization model is also incorporated into the multiple data assimilation iterative ensemble smoothing algorithm, thereby further improving the accuracy of the inversion simulation.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions: The inverse simulation method of CO2 geological storage based on deep learning parameterization strategy includes the following steps: Obtain geological lithology drilling data and hydrogeological data of the target reservoir in the study area, and collect historical observation data of CO2 saturation and reservoir pressure during the storage process at the drilling site; Generate a heterogeneous reservoir permeability field sample using geostatistical software based on the geological lithology drilling data and the hydrogeological data; Constructing a CO2 storage forward numerical model using multiphase flow numerical simulation software based on the geological lithology drilling data and the hydrogeological data; Training a deep learning parameterized model based on the heterogeneous reservoir permeability field samples, and randomly generating an initial sample set of low-dimensional standard normal distribution latent vectors; The deep learning parameterized model, the CO2 storage forward numerical model and the multiple data assimilation iterative ensemble smoothing algorithm jointly construct a CO2 geological storage coupled inversion simulation framework; Based on the initial sample set of low-dimensional standard normal distribution latent vectors, the final low-dimensional latent vector sample set is obtained by using the CO2 geological storage coupled inversion simulation framework; The final low-dimensional latent vector sample set is input into the trained deep learning parameterized model to obtain the corresponding heterogeneous reservoir permeability field posterior sample data set; the heterogeneous reservoir permeability field posterior sample data set is input into the CO2 storage forward numerical model to obtain the corresponding CO2 saturation and reservoir pressure simulation value posterior sample data set.

[0006] Among them, the geostatistical software can be TProGS software. TProGS (Transition Probability Geostatistical Software) is a set of FORTRAN computer programs. It uses the transition probability / Markov chain method to perform geostatistical analysis and simulation on the spatial distribution of geological parameters such as geological units and lithofacies based on known geological drilling data.

[0007] The multiphase flow numerical simulation software can be TOUGH2 (Transport Of Unsaturated Groundwater and Heat) / ECO2N software. TOUGH2 / ECO2N software is the fluid property module of the TOUGH2 simulator. It is specially designed for CO2 geological storage projects in saline aquifers. It can comprehensively and accurately describe the thermodynamic and thermophysical properties of the H2O-NaCl-CO2 mixture.

[0008] Preferably, the deep learning parameterized model construction step includes: Randomly dividing the heterogeneous reservoir permeability field samples into training samples and test samples; During model training, the deep learning parameterized model uses the training samples as input and generates a reconstructed heterogeneous reservoir permeability field from a decoder of the deep learning parameterized model; During model testing, the coefficient of determination between the reconstructed heterogeneous reservoir permeability field generated by the decoder of the deep learning parameterized model trained with the test sample and the training sample is compared. R 2 Precision, judging the coefficient of determination R 2 Whether the preset accuracy requirements are met, if so, it indicates that the deep learning parameterized model training is completed, otherwise continue to use the training samples to optimize and adjust the model parameters until the trained deep learning parameterized model meets the preset accuracy requirements.

[0009] The heterogeneous permeability field is a set of permeability values ​​that change with spatial coordinates in a geological statistical sense at various locations in the reservoir. According to literature research, the heterogeneous permeability field is found to be the most important influencing factor indicator affecting the migration and transformation of CO2 plumes.

[0010] The coefficient of determination R 2 The preset accuracy requirement is the coefficient of determination of the deep learning parameterized model based on the test sample R 2 ≥ α , α is the default value.

[0011] Preferably, the coefficient of determination R 2 : ; In the formula, N is the number of test samples, Represents the test sample, The reconstructed heterogeneous reservoir permeability field generated by the decoder representing the deep learning parameterized model, and represent N The average value of the test samples.

[0012] Preferably, the deep learning parameterized model includes an encoder, a decoder and a discriminator; The encoder is used to extract geostatistical features from the input heterogeneous reservoir permeability field x, and map the geostatistical features to and output a low-dimensional latent vector z; The decoder is used to generate a reconstructed heterogeneous reservoir permeability field from the input low-dimensional latent vector z as well as Standard Deviation As output; The discriminator is used to identify whether the low-dimensional latent vector z obtained by the encoder obeys the standard normal distribution.

[0013] Preferably, the network parameters of the decoder of the deep learning parameterized model are θ As an uncertain parameter, the decoder of the deep learning parameterized model is trained using the SVGD Bayesian inference algorithm to obtain N S Set different network parameters θ The decoder of the deep learning parameterized model of N S Set different network parameters θ The decoder of the deep learning parameterized model is obtained N S Different sets of heterogeneous reservoir permeability fields will be N S The mean value of the permeability field of different heterogeneous reservoirs is used as the output of the decoder of the deep learning parameterized model. ,Will N S The standard deviation of the permeability field of different heterogeneous reservoirs is used as the standard deviation of the permeability field of heterogeneous reservoirs output by the deep learning parameterized model decoder .

[0014] Preferably, the encoder and decoder of the deep learning parameterized model use multi-layer residual dense blocks as basic modules, each of which is composed of 3 residual dense blocks, each of which contains 5 internal convolutional layers, and each of which includes batch normalization, ReLU nonlinear activation function and convolution operation; the multi-layer residual dense block makes full use of its dense connection and residual learning characteristics, which can increase the propagation of information flow within the network, alleviate the burden of training complex networks, and thus improve the deep learning parameterized model's characterization of the geological statistical characteristics of complex heterogeneous permeability fields.

[0015] The discriminator of the deep learning parameterized model includes a first convolutional layer, a second convolutional layer, a first fully connected layer, and a second fully connected layer. The first convolutional layer, the second convolutional layer, the first fully connected layer, and the second fully connected layer are connected in sequence, and the second fully connected layer uses a Sigmoid function as a nonlinear activation function to ensure that a value between 0 and 1 is output, indicating the probability that the low-dimensional latent vector z obtained by the encoder of the deep learning parameterized model follows a standard normal distribution.

[0016] Preferably, the loss function used by the encoder and decoder of the deep learning parameterized model is for: ; In the formula, Represents the encoder i Input fields, Represents the first i Output fields, represents the encoder, represents the discriminator, N is the number of samples, w is the weight factor; The loss function used by the discriminator of the deep learning parameterized model is for: ; In the formula, Representative i A low-dimensional normally distributed latent vector.

[0017] Among them, training the encoder, decoder and discriminator of the deep learning parameterized model is an adversarial training process: the encoder constantly deceives the discriminator so that the encoding vector output by the encoder The probability of following the standard normal distribution is higher, and the discriminator constantly distinguishes whether the input vector is a real sample from the standard normal distribution or a "fake" sample generated by the encoder. After this adversarial training, when the input arbitrary low-dimensional latent vector z is a normal distribution, the decoder is able to generate a reconstructed sample with similar geostatistical characteristics to the input field x. as well as Standard Deviation .

[0018] Preferably, the CO2 geological storage coupled inversion simulation framework includes: A decoder of a deep learning parameterized model, wherein the decoder of the deep learning parameterized model takes a low-dimensional standard normal distribution latent vector sample set as input and outputs a heterogeneous reservoir permeability field sample set and a heterogeneous reservoir permeability field standard deviation sample set; A CO2 storage forward numerical model, wherein the CO2 storage forward numerical model takes a heterogeneous reservoir permeability field sample set output by a decoder of a deep learning parameterized model as an input, and outputs a CO2 saturation and reservoir pressure simulation value sample set; A multiple data assimilation iterative ensemble smoothing algorithm, which uses the low-dimensional standard normal distribution latent vector sample set, the output of the CO2 storage forward numerical model, the historical observation data of CO2 saturation and reservoir pressure, and the standard deviation sample set of the heterogeneous reservoir permeability field as input data of the multiple data assimilation iterative ensemble smoothing algorithm, and outputs an updated low-dimensional standard normal distribution latent vector sample set.

[0019] Preferably, the CO2 geological storage coupled inversion simulation framework execution steps include: S1, randomly generate a series of low-dimensional standard normal distribution latent vector initial sample sets; S2, inputting a series of randomly generated low-dimensional standard normal distribution latent vector initial sample sets into the decoder of the trained deep learning parameterized model to obtain the corresponding heterogeneous reservoir permeability field initial sample set and heterogeneous reservoir permeability field standard deviation initial sample set; S3, inputting the initial sample set of the heterogeneous reservoir permeability field into the CO2 storage forward numerical model to obtain the corresponding initial sample set of CO2 saturation and reservoir pressure simulation values; S4, based on the collected historical observation data of CO2 saturation and reservoir pressure, the initial sample set of standard deviation of the heterogeneous reservoir permeability field and the initial sample set of simulated values ​​of CO2 saturation and reservoir pressure, using the multiple data assimilation iterative set smoothing algorithm to update the low-dimensional latent vector sample set; S5. Determine whether the preset maximum number of cycles has been reached. If the preset maximum number of cycles has not been reached, input the low-dimensional latent vector sample set updated by the multiple data assimilation iterative set smoothing algorithm into the decoder of the trained deep learning parameterized model to obtain the corresponding heterogeneous reservoir permeability field sample set and the heterogeneous reservoir permeability field standard deviation sample set, and jump to step S3 to continue executing until the preset maximum number of cycles has been reached to obtain the final low-dimensional latent vector sample set.

[0020] Preferably, the multiple data assimilation iterative ensemble smoothing algorithm: ; In the formula, represents the set of low-dimensional latent vector samples before updating, represents the updated low-dimensional latent vector sample set, is a sample set of simulated values ​​of CO2 saturation and reservoir pressure, is the cross-covariance matrix between the low-dimensional latent vector sample set and the CO2 saturation and reservoir pressure simulation value sample set, represents the autocovariance matrix of the sample set of simulated values ​​of CO2 saturation and reservoir pressure, is the historical observation data, is the autocovariance matrix of the standard deviation of the historical observation data error, E is the autocovariance matrix of the standard deviation of the heterogeneous reservoir permeability field output by the deep learning parameterized model decoder; ; ;in, is the set size of the low-dimensional latent vector sample set, The preset maximum number of cycles for the multiple data assimilation iterative ensemble smoothing algorithm; is the coefficient of expansion and α = N iter .

[0021] Compared with the prior art, the beneficial effects achieved by the present invention are: 1. The present invention effectively maps the non-Gaussian distribution characteristics of the reservoir permeability field to the latent vector space of the low-dimensional normal distribution, greatly reducing the number of parameters to be estimated in the CO2 geological storage inversion simulation, and can estimate the parameter field of the non-Gaussian distribution.

[0022] 2. By constructing the encoder, decoder and discriminator components in the deep learning parameterized model, and using multi-layer residual dense blocks with dense connections and residual learning characteristics as basic modules, the model can enhance the characterization of the geological statistical characteristics of complex heterogeneous permeability fields.

[0023] 3. The decoder of the trained deep learning parameterized model takes any low-dimensional standard normal distribution latent vector as input and can output the corresponding heterogeneous reservoir permeability field and the standard deviation field of the heterogeneous reservoir permeability field, and the heterogeneous reservoir permeability field retains its non-Gaussian distribution geological statistical characteristics.

[0024] 4. The decoder of the deep learning parameterized model, the forward numerical model of CO2 storage and the multiple data assimilation iterative ensemble smoothing algorithm jointly construct a coupled inversion simulation framework for CO2 geological storage, which can effectively reproduce the historical process of CO2 plume migration in the reservoir. During the inversion, only the low-dimensional latent vector needs to be updated cyclically to efficiently estimate the heterogeneous permeability field of complex reservoirs, thereby improving the simulation and prediction accuracy of the numerical model for the CO2 geological storage process, and can be used to predict the possible migration law of CO2 in the future.

[0025] 5. In the multiple data assimilation iterative ensemble smoothing algorithm, the standard deviation of the heterogeneous reservoir permeability field output by the deep learning parameterized model decoder is also incorporated into the algorithm formula as an error term, thereby further improving the accuracy of the inversion simulation. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1 It is a flowchart of the steps of obtaining the final low-dimensional standard normal distribution latent vector sample set by using the CO2 geological storage coupled inversion simulation framework of the present invention; Figure 2 It is a schematic diagram of the deep learning parameterization (DLP) model structure; Figure 3 It is a schematic diagram of the structure of the multi-layer residual dense block (MLRDB), which is the basic component module of the encoder and decoder of the deep learning parameterized model; Figure 4 It is a sample comparison diagram of three groups of real permeability fields, the reconstructed permeability fields generated by the decoder of the deep learning parameterized model, and the standard deviation field of the reconstructed permeability fields; Figure 5 is the heterogeneous permeability of the CO2 storage forward numerical model (log 10 k ) Reference field diagram and permeability histogram; Figure 6 This is the inversion result diagram of the heterogeneous permeability field; Figure 7 It is the inversion result diagram of CO2 saturation field at four moments; Figure 8 It is the inversion result diagram of reservoir pressure increment field at four moments. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] The CO2 geological storage site selected by the present invention is the Frio carbon storage site in the southwestern United States. It is a typical fluvial sedimentary reservoir with a high degree of channelization. The lithology is mainly composed of high-permeability sandstone and low-permeability shale. The present invention is further described, in which the deep learning parameterized model can be referred to as the DLP model.

[0029] The present invention provides a technical solution, a CO2 geological storage inversion simulation method based on a deep learning parameterization strategy, and its specific execution steps include: L1. The study area is determined to be the Frio storage site in the southwestern United States. The target reservoir is a fluvial sedimentary reservoir with a highly channelized lithofacies composition. Geological drilling shows that sandstone accounts for about 75% of the target reservoir with an average permeability of 4×10 -13 m 2 , porosity 0.30, capillary pressure 20 kPa, pore distribution index 0.457; shale accounts for about 25%, with an average permeability of 1×10 -18 m 2 , porosity 0.10, capillary pressure 20000kPa, pore distribution index 0.457; reservoir burial depth 1500m, initial average pressure 15MPa, temperature 60℃ (isothermal), salinity 0.03, rock anisotropy ratio 0.5. Historical observation data of CO2 saturation and reservoir pressure were collected at the drilling site, and Gaussian noise with a standard deviation of 2% was added as the observation error of the historical observation data.

[0030] L2. Generate 40,000 sets of heterogeneous reservoir permeability field samples that conform to drilling and hydrogeological data through the geological statistics software TProGS, of which 35,000 sets are used as training samples to optimize the training DLP model, and 5,000 sets are used as test samples to evaluate the test accuracy of the DLP model.

[0031] Among them, the DLP model is as follows Figure 2 As shown in Figure 1, it consists of an encoder, a decoder, and a discriminator. The basic building block of the encoder and decoder is the multi-layer residual dense block (MLRDB), whose structure is as follows: Figure 3As shown in the figure, each MLRDB consists of 3 residual dense blocks (RDBs), each RDB contains 5 internal convolutional layers, and each convolutional layer includes batch normalization (BN), ReLU nonlinear activation function and convolution (Conv) operation. N f Represents the number of feature maps, and the residual scaling factor is set to β = 0.2, ⊕ represents the addition operation.

[0032] Figure 2 The encoder of the DLP model also involves a downsampling operation to fully extract features from the heterogeneous permeability field x and map it to a low-dimensional latent vector z. The decoder also involves an upsampling operation to restore and reconstruct the heterogeneous permeability field from the low-dimensional latent vector z. The reshape operation is used in the discriminator, and the last two layers are fully connected (FC) layers with 128 and 1 neurons respectively. The last fully connected layer uses a Sigmoid activation function to output a value between 0 and 1, indicating the probability that the low-dimensional latent vector z obtained by the encoder follows the standard normal distribution.

[0033] L3, the DLP model was trained using 35,000 sets of training samples. The decoder was trained using the SVGD Bayesian inference algorithm with an initial learning rate of 0.0025, and the encoder and discriminator were trained using the Adam optimizer with an initial learning rate of 0.0002. The sample batch size was set to 16, the number of training rounds was set to 50, and the loss function weight factor used by the encoder and decoder was w =0.01.

[0034] When training the decoder, the SVGD Bayesian inference algorithm uses the decoder's network parameters θ As an uncertain parameter, we get N S = 20 different network parameters θ The low-dimensional latent vector z is input to N S = 20 different network parameters θ The corresponding decoder obtains N S = 20 sets of different heterogeneous reservoir permeability field data, N S = The average value of 20 sets of heterogeneous reservoir permeability field data is used as the heterogeneous reservoir permeability field output by the decoder ,Will N S = The standard deviation of 20 sets of heterogeneous reservoir permeability field data as the standard deviation of the heterogeneous reservoir permeability field output by the decoder .

[0035] The test accuracy of the DLP model is evaluated based on 5000 test samples, and the determination coefficient ( R 2 ) precision, and R 2 Optimize and adjust the trained DLP model for the discrimination criteria until R 2 Meet the preset accuracy requirements (determination coefficient R 2 ≥ α , α is the preset value), and the final trained DLP model is obtained.

[0036] The final calculated coefficient of determination R 2 It is 0.961, indicating that the DLP model has met the accuracy requirements and can effectively explore the geostatistical characteristics of the heterogeneous permeability field, map it to a low-dimensional latent vector space, and reconstruct the permeability field.

[0037] Figure 4 The three groups of real permeability fields (randomly selected from 5000 test samples) were compared with the reconstructed permeability fields and the standard deviation fields of the reconstructed permeability fields generated by the decoders of the corresponding DLP models. As shown in the figure, the real fields and reconstructed fields of the three groups of samples are very close, indicating that the DLP model can accurately approximate complex heterogeneous permeability fields. The standard deviation of the reconstructed permeability field is also basically below 0.5, indicating that the uncertainty of the output results of the DLP model decoder is small, and the inversion simulation results obtained by incorporating this error term into the multiple data assimilation iterative ensemble smoothing algorithm are more reliable.

[0038] At this time, the decoder of the DLP model is incorporated into the CO2 geological storage coupled inversion simulation framework. Given any low-dimensional standard normal distribution latent vector as input, the corresponding heterogeneous reservoir permeability field and heterogeneous reservoir permeability standard deviation field can be output.

[0039] L4. The forward numerical model of CO2 storage was established using the multiphase flow simulation software TOUGH2 / ECO2N. The simulated storage area size is 7200 m×1000 m×100 m. The reservoir is discretized into 35×25×10 grids. The CO2 injection well (well 5) is located in the center of the reservoir, and the grid is appropriately encrypted around the CO2 injection well. CO2 is continuously injected at an injection rate of 2.0 kg / s for 10 years. The heterogeneous permeability (log 10 k ) Reference field and 9 well locations are as follows Figure 5As shown in a, Figure 5 The b in FIG. 5 is a histogram of permeability, and its distribution shows obvious non-Gaussian characteristics.

[0040] L5, the decoder in the trained DLP model, the CO2 storage forward numerical model and the iterative ensemble smoothing of multiple data assimilation (ESMDA) algorithm are used to jointly construct the CO2 geological storage coupled inversion simulation framework; Among them, the CO2 geological storage coupled inversion simulation framework includes: an encoder of a deep learning parameterized model, which takes a low-dimensional standard normal distribution latent vector sample set as input and outputs a heterogeneous reservoir permeability field sample set and a heterogeneous reservoir permeability field standard deviation sample set; The CO2 storage forward numerical model takes the heterogeneous reservoir permeability field sample set output by the decoder of the deep learning parameterized model as input, and outputs the CO2 saturation and reservoir pressure simulation value sample set; The multiple data assimilation iterative ensemble smoothing algorithm takes a low-dimensional standard normal distribution latent vector sample set, a CO2 saturation and reservoir pressure simulation value sample set, a CO2 saturation and reservoir pressure historical observation data and a heterogeneous reservoir permeability field standard deviation sample set as input, and outputs an updated low-dimensional standard normal distribution latent vector sample set.

[0041] L6. Randomly generate 500 sets of low-dimensional latent vector initial sample sets, and use the CO2 geological storage coupled inversion simulation framework to obtain the final low-dimensional latent vector sample set. The specific steps are: The 500 sets of initial sample sets of low-dimensional latent vectors are input into the decoder of the DLP model trained in L3 to obtain 500 sets of initial sample sets of heterogeneous reservoir permeability fields and initial sample sets of standard deviations of heterogeneous reservoir permeability fields.

[0042] An initial sample set of 500 sets of heterogeneous reservoir permeability fields was input into the L4 CO2 storage forward numerical model to obtain an initial sample set of 500 sets of CO2 saturation and reservoir pressure simulation values.

[0043] Based on the historical observation data of CO2 saturation and reservoir pressure collected in L1, 500 sets of initial sample sets of standard deviation of heterogeneous reservoir permeability field and 500 sets of initial sample sets of CO2 saturation and reservoir pressure simulation values, 500 sets of initial sample sets of low-dimensional latent vectors are updated through the ESMDA algorithm. The above process is repeated until the cycle is repeated 30 times to obtain the final low-dimensional latent vector sample set. The set size of the low-dimensional latent vector sample set in the ESMDA algorithm is The preset maximum number of cycles is 500. is 30.

[0044] L7, the low-dimensional latent vector sample set obtained in the final (30th update) is input into the decoder of the DLP model trained in L3 to obtain the corresponding heterogeneous permeability field posterior sample data set. The results are as follows Figure 6 shown.

[0045] in, Figure 6 The inversion result diagram of the heterogeneous permeability field is shown, including the true permeability field, the posterior permeability mean field and the posterior permeability standard deviation field. The locations of the 9 wells are still marked in the figure.

[0046] The inversion results of the heterogeneous permeability field show that the proposed CO2 geological storage coupled inversion simulation framework successfully estimates and inverts the general distribution characteristics of the real permeability field, and the values ​​of the posterior standard deviation field are relatively small in areas with observation wells, indicating that the accuracy and stability of the inversion simulation are relatively high.

[0047] L8, the heterogeneous permeability field posterior sample data set is input into the CO2 storage forward numerical model in L4 to obtain the CO2 saturation and reservoir pressure simulation value posterior sample data set. The results are as follows Figure 7 and 8 shown.

[0048] in, Figure 7 The inversion results of the CO2 saturation field at four moments (2.5 years, 5 years, 7.5 years, and 10 years) are shown, including the true field of CO2 saturation, the posterior mean field of CO2 saturation, and the posterior standard deviation field of CO2 saturation. The locations of the nine wells are still marked in the figure.

[0049] Figure 8 The inversion results of the reservoir pressure increment field at four moments (2.5 years, 5 years, 7.5 years, and 10 years) are shown, including the true field of reservoir pressure increment, the posterior mean field of reservoir pressure increment, and the posterior standard deviation field of reservoir pressure increment. The locations of the nine wells are still marked in the figure.

[0050] Depend on Figure 7 and Figure 8 It can be seen that the CO2 geological storage inversion simulation method based on deep learning parameterization strategy proposed in the present invention can achieve good inversion simulation results. At any time, the posterior mean value field obtained by inversion is very close to the real field in the area covered by observation wells in the middle of the reservoir. In areas without observation wells, the similarity between the two decreases with the increase of the distance from the observation well. And the values ​​of the posterior standard deviation field are relatively small, indicating that the obtained posterior sample data set has high stability, which can effectively characterize the historical process of CO2 plume migration in complex reservoirs, and can provide technical support for the optimized operation and risk management of CO2 geological storage projects.

[0051] It is worth mentioning that in this implementation case, the deep learning parameterization model reduced the dimension of a complex 8750-dimensional heterogeneous permeability field to a low-dimensional latent vector of 378 dimensions, which increased the efficiency of the multiple data assimilation iterative ensemble smoothing algorithm in estimating parameters by about 23 times, and successfully accelerated the calculation process of CO2 geological storage inversion simulation. In addition, the present invention is not limited to the dimensionality reduction of high-dimensional heterogeneous permeability fields for CO2 geological storage numerical simulations, but also has broad application prospects for the dimensionality reduction of parameter fields for other similar multiphase flow numerical simulations (such as DNAPL solute migration simulations, etc.).

[0052] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0053] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A CO2 geological storage inversion simulation method based on deep learning parameterization strategy, characterized in that: The steps include: Obtain geological lithology drilling data and hydrogeological data of the target reservoir in the study area, and collect historical observation data of CO2 saturation and reservoir pressure during the storage process at the drilling site; Generate a heterogeneous reservoir permeability field sample using geostatistical software based on the geological lithology drilling data and the hydrogeological data; Constructing a CO2 storage forward numerical model using multiphase flow numerical simulation software based on the geological lithology drilling data and the hydrogeological data; Training a deep learning parameterized model based on the heterogeneous reservoir permeability field samples, and randomly generating an initial sample set of low-dimensional standard normal distribution latent vectors; The deep learning parameterized model, the CO2 storage forward numerical model and the multiple data assimilation iterative ensemble smoothing algorithm jointly construct a CO2 geological storage coupled inversion simulation framework; Based on the initial sample set of low-dimensional standard normal distribution latent vectors, the final low-dimensional latent vector sample set is obtained by using the CO2 geological storage coupled inversion simulation framework; The final low-dimensional latent vector sample set is input into the trained deep learning parameterized model to obtain the corresponding heterogeneous reservoir permeability field posterior sample data set; the heterogeneous reservoir permeability field posterior sample data set is input into the CO2 storage forward numerical model to obtain the corresponding CO2 saturation and reservoir pressure simulation value posterior sample data set.

2. The CO2 geological storage inversion simulation method based on deep learning parameterization strategy according to claim 1 is characterized in that: The deep learning parameterized model construction step includes: Randomly dividing the heterogeneous reservoir permeability field samples into training samples and test samples; During model training, the deep learning parameterized model uses the training samples as input and generates a reconstructed heterogeneous reservoir permeability field from a decoder of the deep learning parameterized model; During model testing, the coefficient of determination between the reconstructed heterogeneous reservoir permeability field generated by the decoder of the deep learning parameterized model trained with the test sample and the training sample is compared. R 2 Precision, judging the coefficient of determination R 2 Whether the preset accuracy requirements are met, if so, it indicates that the deep learning parameterized model training is completed, otherwise continue to use the training samples to optimize and adjust the model parameters until the trained deep learning parameterized model meets the preset accuracy requirements.

3. The CO2 geological storage inversion simulation method based on deep learning parameterization strategy according to claim 2 is characterized in that: The coefficient of determination R 2 : ； In the formula, N is the number of test samples, Represents the test sample, The reconstructed heterogeneous reservoir permeability field generated by the decoder representing the deep learning parameterized model, and represent N The average value of the test samples.

4. The CO2 geological storage inversion simulation method based on deep learning parameterization strategy according to claim 1 is characterized in that: The deep learning parameterized model includes an encoder, a decoder, and a discriminator; The encoder is used to extract geostatistical features from the input heterogeneous reservoir permeability field x, and map the geostatistical features to and output a low-dimensional latent vector z; The decoder is used to generate a reconstructed heterogeneous reservoir permeability field from the input low-dimensional latent vector z as well as Standard Deviation As output; The discriminator is used to identify whether the low-dimensional latent vector z obtained by the encoder obeys the standard normal distribution.

5. The CO2 geological storage inversion simulation method based on deep learning parameterization strategy according to claim 4 is characterized in that: The network parameters of the decoder of the deep learning parameterized model θ As an uncertain parameter, the decoder of the deep learning parameterized model is trained using the SVGD Bayesian inference algorithm to obtain N S Set different network parameters θ The decoder of the deep learning parameterized model of N S Set different network parameters θ The decoder of the deep learning parameterized model is obtained N S Different sets of heterogeneous reservoir permeability fields will be N S The mean value of the permeability field of different heterogeneous reservoirs is used as the output of the decoder of the deep learning parameterized model. ,Will N S The standard deviation of the permeability field of different heterogeneous reservoirs is used as the standard deviation of the permeability field of heterogeneous reservoirs output by the deep learning parameterized model decoder .

6. The CO2 geological storage inversion simulation method based on deep learning parameterization strategy according to claim 4 is characterized in that: The encoder and decoder of the deep learning parameterized model use multi-layer residual dense blocks as basic modules, each of which is composed of 3 residual dense blocks, each of which contains 5 internal convolutional layers, and each of which includes batch normalization, ReLU nonlinear activation function and convolution operation; The discriminator of the deep learning parameterized model includes a first convolutional layer, a second convolutional layer, a first fully connected layer, and a second fully connected layer; the first convolutional layer, the second convolutional layer, the first fully connected layer, and the second fully connected layer are connected in sequence, and the second fully connected layer uses a Sigmoid function as a nonlinear activation function to ensure that a value between 0 and 1 is output, indicating the probability that the low-dimensional latent vector z obtained by the encoder of the deep learning parameterized model follows a standard normal distribution.

7. The CO2 geological storage inversion simulation method based on deep learning parameterization strategy according to claim 4 is characterized in that: The loss function used by the encoder and decoder of the deep learning parameterized model for: ； In the formula, Represents the encoder i Input fields, Represents the first i Output fields, represents the encoder, represents the discriminator, N is the number of samples, w is the weight factor; The loss function used by the discriminator of the deep learning parameterized model is for: ； In the formula, Representative i A low-dimensional normally distributed latent vector.

8. The CO2 geological storage inversion simulation method based on deep learning parameterization strategy according to claim 1, characterized in that: The CO2 geological storage coupled inversion simulation framework includes: A decoder of a deep learning parameterized model, wherein the decoder of the deep learning parameterized model takes a low-dimensional standard normal distribution latent vector sample set as input and outputs a heterogeneous reservoir permeability field sample set and a heterogeneous reservoir permeability field standard deviation sample set; A CO2 storage forward numerical model, wherein the CO2 storage forward numerical model takes a heterogeneous reservoir permeability field sample set output by a decoder of a deep learning parameterized model as an input, and outputs a CO2 saturation and reservoir pressure simulation value sample set; A multiple data assimilation iterative ensemble smoothing algorithm, which uses the low-dimensional standard normal distribution latent vector sample set, the output of the CO2 storage forward numerical model, the historical observation data of CO2 saturation and reservoir pressure, and the standard deviation sample set of the heterogeneous reservoir permeability field as input data of the multiple data assimilation iterative ensemble smoothing algorithm, and outputs an updated low-dimensional standard normal distribution latent vector sample set.

9. The CO2 geological storage inversion simulation method based on deep learning parameterization strategy according to claim 1, characterized in that: The steps of executing the CO2 geological storage coupled inversion simulation framework include: S1, randomly generate a series of low-dimensional standard normal distribution latent vector initial sample sets; S2, inputting a series of randomly generated low-dimensional standard normal distribution latent vector initial sample sets into the decoder of the trained deep learning parameterized model to obtain the corresponding heterogeneous reservoir permeability field initial sample set and heterogeneous reservoir permeability field standard deviation initial sample set; S3, inputting the initial sample set of the heterogeneous reservoir permeability field into the CO2 storage forward numerical model to obtain the corresponding initial sample set of CO2 saturation and reservoir pressure simulation values; S4, based on the collected historical observation data of CO2 saturation and reservoir pressure, the initial sample set of standard deviation of the heterogeneous reservoir permeability field and the initial sample set of simulated values ​​of CO2 saturation and reservoir pressure, using the multiple data assimilation iterative set smoothing algorithm to update the low-dimensional latent vector sample set; S5. Determine whether the preset maximum number of cycles has been reached. If the preset maximum number of cycles has not been reached, input the low-dimensional latent vector sample set updated by the multiple data assimilation iterative set smoothing algorithm into the decoder of the trained deep learning parameterized model to obtain the corresponding heterogeneous reservoir permeability field sample set and the heterogeneous reservoir permeability field standard deviation sample set, and jump to step S3 to continue executing until the preset maximum number of cycles has been reached to obtain the final low-dimensional latent vector sample set.

10. The CO2 geological storage inversion simulation method based on deep learning parameterization strategy according to claim 1, characterized in that: The multiple data assimilation iterative ensemble smoothing algorithm: ； In the formula, represents the set of low-dimensional latent vector samples before updating, represents the updated low-dimensional latent vector sample set, is a sample set of simulated values ​​of CO2 saturation and reservoir pressure, is the cross-covariance matrix between the low-dimensional latent vector sample set and the CO2 saturation and reservoir pressure simulation value sample set, represents the autocovariance matrix of the sample set of simulated values ​​of CO2 saturation and reservoir pressure, is the historical observation data, is the autocovariance matrix of the standard deviation of the historical observation data error, E is the autocovariance matrix of the standard deviation of the heterogeneous reservoir permeability field output by the deep learning parameterized model decoder; ; ;in, is the set size of the low-dimensional latent vector sample set, The preset maximum number of cycles for the multiple data assimilation iterative ensemble smoothing algorithm; is the coefficient of expansion and α = N iter .

Images